From 780623261eedd996404795dfb7928e680408adb5 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 11 Jan 2016 09:07:14 -0800
Subject: Re-enabled the optimized reduction CUDA code.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h     | 2 --
 unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h | 6 +++---
 2 files changed, 3 insertions(+), 5 deletions(-)

(limited to 'unsupported/Eigen/CXX11/src/Tensor')

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index fd7064459..cea32d05f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -527,7 +527,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
     }
 
     // Attempt to use an optimized reduction.
-#if 0
     else if (RunningOnGPU && data && (m_device.majorDeviceVersion() >= 3)) {
       bool reducing_inner_dims = true;
       for (int i = 0; i < NumReducedDims; ++i) {
@@ -563,7 +562,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
         return false;
       }
     }
-#endif
     return true;
   }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index 558d0c83d..198b3604c 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -126,7 +126,7 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
     const int block_size = 256;
     const int num_per_thread = 128;
     const int num_blocks = std::ceil(static_cast<float>(num_coeffs) / (block_size * num_per_thread));
-    LAUNCH_CUDA_KERNEL((FullReductionKernel<block_size, num_per_thread>),
+    LAUNCH_CUDA_KERNEL((FullReductionKernel<block_size, num_per_thread, Self, Op, Index>),
                        num_blocks, block_size, 0, device, reducer, self, num_coeffs, output);
   }
 };
@@ -222,7 +222,7 @@ struct InnerReducer<Self, Op, GpuDevice> {
     const int num_per_thread = 128;
     const int num_blocks = 32;
 
-    LAUNCH_CUDA_KERNEL((InnerReductionKernel<num_per_thread>),
+    LAUNCH_CUDA_KERNEL((InnerReductionKernel<num_per_thread, Self, Op, Index>),
                        num_blocks, block_size, block_size*sizeof(float), device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
   }
 };
@@ -279,7 +279,7 @@ struct OuterReducer<Self, Op, GpuDevice> {
                            device.maxCudaThreadsPerMultiProcessor() / block_size;
     const int num_blocks = numext::mini<int>(max_blocks, dyn_blocks);
 
-    LAUNCH_CUDA_KERNEL((OuterReductionKernel<num_per_thread>),
+    LAUNCH_CUDA_KERNEL((OuterReductionKernel<num_per_thread, Self, Op, Index>),
                        num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
   }
 };
-- 
cgit v1.2.3